Two ways to make a language model know maritime: put the knowledge in the weights (fine-tuning) or put it in a library the model searches at query time (retrieval-augmented generation, RAG). For most domains the choice is a preference; for maritime it's a consequence of the domain's structure — a corpus of rules that changes between model generations, and answers that need to be citable to a source.
The honest answer is not 'RAG or fine-tuning' — it's which task needs which property, and why the safety-critical ones need both.
Key findings
- Maritime knowledge changes at a rate that makes weight-frozen knowledge go stale between model versions — retrieval solves the freshness problem directly. Established
- RAG's unique maritime value is the citation: an answer that must point to the current MARPOL annex or MASS circular needs to retrieve, not memorise. Established
- Fine-tuning is the better tool for <em>behaviour and format</em> — making a model reliably produce structured COLREGS-reasoning output is a fine-tuning problem. Research demonstration
- The strongest published maritime-LLM work (OceanGPT, agentic twins, diagnostic agents) uses hybrids — retrieval for facts, weights for style and skill. Established
- Retrieval failure is the RAG risk aboard: a query that surfaces the wrong passage produces a confidently wrong answer that cites a real source. Research demonstration
The freshness argument for retrieval
Maritime regulation moves. IMO circulars issue; MARPOL annexes amend; MASS guidance evolves; flag notices change. A fine-tuned model's knowledge is frozen at training time — an answer about 'current' ECA boundaries or MASS requirements is only as current as the last training run. Retrieval sidesteps this entirely: the model fetches the current document at query time and answers from it. For regulatory and compliance queries, retrieval isn't an optimisation — it's the only architecture that stays correct. Established
The skill argument for fine-tuning
Fine-tuning's advantage isn't knowledge — it's behaviour. A model trained on maritime reasoning produces the domain's structures natively: the COLREGS-encounter framing, the machinery-diagnostic workflow, the route-planning decomposition. This is what the specialised small-model work shows: a 14B model fine-tuned on COLREGS outperforms a 72B general model not because it knows more, but because it has learned to reason in the domain's form. Retrieval can't teach a model a skill; fine-tuning can't keep it current. Research demonstration
| Need | RAG | Fine-tuning | Right choice |
|---|---|---|---|
| Current regulatory facts | Yes — fetches live docs | No — frozen at train time | RAG |
| Structured reasoning style | Weak — format learned better | Yes — trained behaviour | Fine-tuning |
| Citable / auditable answers | Yes — returns the source | No — no source | RAG |
| Offline / edge inference | Needs the corpus local | Self-contained weights | Fine-tuning + local corpus |
| Rare-domain knowledge | Retrieves if indexed | Needs it in training data | RAG for corpus'd topics |
The maritime corpus problem
RAG is only as good as its index. The maritime corpus — IMO documents, flag circulars, classification rules, nautical publications, the vessel's own manuals — is heterogeneous, versioned, and partially paywalled. Building the retrieval index is the unglamorous work that decides whether the RAG answer is right. A model with a stale or incomplete index confidently cites last year's rule. The corpus is the system; the model is the interface to it. Maritime-GPT analysis
The hybrid the evidence supports
The consistent pattern in the strongest maritime LLM work is hybrid: fine-tune the model into the domain's reasoning style, then ground its factual answers in retrieval. OceanGPT builds domain instruction data for the skill; the agentic-diagnostics work retrieves sensor readings rather than memorising them; the tool-calling pattern is retrieval generalised to live systems. The answer is neither pole — it's a model fluent in the domain's form, grounded in the domain's current documents. Established
What we know
- Regulatory freshness and citability require retrieval; reasoning style requires fine-tuning.
- The strongest maritime-LLM work combines both.
- The retrieval index quality is the RAG system's real vulnerability.
What we don't
- A published benchmark of RAG vs fine-tuned maritime-LLM accuracy on the same tasks.
- How retrieval failure rates behave aboard with a partial or stale index.
- Whether small edge-deployable models can run effective local RAG.
Being tested
- Maritime-specific retrieval corpora and benchmarks.
- Hybrid FT+RAG architectures for navigation and compliance queries.
- Offline RAG — a compressed regulatory corpus on the ship's own hardware.
Maritime-GPT analysis
The RAG-vs-fine-tuning question dissolves once you stop asking it as a binary. Retrieval is the freshness-and-citation layer; fine-tuning is the skill-and-style layer. Maritime needs both because maritime is both — a domain of rules that change and reasoning that must be structured. The systems that will work aboard are the hybrids, and the differentiator will be the quality of the maritime corpus they retrieve from, not the model name. Maritime-GPT analysis
This is a research feature, not operational guidance. Simulation results are reported as simulation; demonstrations are not deployments; company claims are attributed as such. Verify safety-critical decisions against official nautical publications and flag-state requirements.
Sources & evidence
The foundational RAG paper — the architecture the maritime application inherits.
The tool-calling/retrieval pattern in a validated maritime diagnostic role.
The edge-compute constraint that shapes the RAG/FT decision aboard.